Since the brain operates as a scale-free network, the frequency-power plot of fMRI BOLD signals follows a power-law distribution. However, conventional deep neural network models that fail to account for this characteristic tend to underperform, lacking a proper model structure to capture the brain signal as a complex system adequately. In this paper, we propose DivfreqBERT, an end-to-end timeseries data model that leverages the features of scale-free networks to encode biological characteristics better. We employed Lorentzian and multi-fractal functions to partition the whole-brain dynamics into three components, each consistently following the power-law function while exhibiting distinct characteristics of small-worldness in spatial connectivity. Merely adjusting the encoding method resulted in enhanced performance in sex classification for both the ABCD (Adolescent Brain Cognitive Development) and UKB (UK Biobank) datasets. Additionally, in the pretraining phase, we utilized the differences in small-worldness across different frequency ranges to mask nodes in order of their communicability, allowing the model to learn the structure where highly communicable nodes influence other nodes. After pre-training the model on UKB data, fine-tuning it on ABCD data yielded remarkable performance improvements. DivfreqBERT demonstrates the utility of knowledge-guided deep neural networks by leveraging complex system properties.